Your RTX Pro 6000 Blackwell Does Not Support FlashAttention-4

👁 rtx pro 6000 blackwell flashattention 4 support

If you bought an RTX Pro 6000 Blackwell expecting full Blackwell support for local LLM inference, you will not get FlashAttention-4. That kernel only runs on datacenter Blackwell GPUs like NVIDIA B200 and on NVIDIA H100.

Even though the branding says “Blackwell”, the underlying hardware is different in a way that directly affects inference performance.

Why FlashAttention-4 matters for inference

FlashAttention-4 is not a small upgrade. On B200 it pushes attention close to pure matmul speed, hitting around 1.6 PFLOPs/s in BF16 forward passes and showing 2x+ gains over Triton kernels. In practical terms, this reduces one of the main bottlenecks in long context inference.

For local LLM users running vLLM or PyTorch FlexAttention, this means higher tokens per second, especially on large models with long KV cache usage.

But these gains depend on very specific hardware features.

The real issue: SM versions, not branding

The key concept here is SM, or Streaming Multiprocessor. This is the core compute unit inside NVIDIA GPUs. Each SM version defines what instructions and hardware features are available.

There are effectively two different Blackwell families:

Datacenter Blackwell uses SM100.
Consumer and workstation Blackwell uses SM120 or SM121.

The difference is not cosmetic. It changes what kernels can run.

SM100 GPUs like B200 include:

• Dedicated tensor memory (TMEM)
• New tensor instructions (tcgen05)
• Advanced scheduling features for matrix pipelines

SM120 GPUs like the RTX Pro 6000:

• No TMEM
• No tcgen05
• Fallback to extended mma.sync, similar to Ampere

FlashAttention-4 is built specifically around SM100 features like TMEM and async tensor pipelines. Without those, the kernel cannot run at all, not even in a degraded mode.

What happens on RTX Pro 6000

On SM120 hardware, frameworks fall back to older paths:

Triton kernels behave closer to Ampere
cuDNN attention or FA-2/FA-3 style kernels are used instead
Some frameworks treat the GPU as a legacy architecture for compatibility

This means your expensive workstation GPU can end up running attention similarly to a much older card like an RTX 3090 in terms of kernel design, even if raw compute is higher.

Practical impact for local LLM setups

For most local inference builds, the bottleneck is still VRAM and memory bandwidth. In that sense, the RTX Pro 6000 is still useful if you need large memory capacity.

But if your goal is maximum tokens per second per watt or per dollar, the lack of FA-4 support matters.

You are missing:

• The latest attention optimizations
• Better scaling on long context
• Future kernels targeting SM100 first

This creates a gap where datacenter GPUs pull further ahead in efficiency, not just raw power.

The uncomfortable comparison

A single NVIDIA B200 can fully use FA-4 and newer kernels. But it requires HGX systems and extreme cost.

Meanwhile, RTX Pro 6000 fits in a workstation and is far more accessible, but lacks the instruction set needed for those kernels.

For price conscious builders, this creates an awkward middle ground. You pay for “Blackwell”, but you get a software experience closer to older architectures.

What to expect going forward

The good news is that algorithmic ideas from FlashAttention-4, like selective rescaling and improved softmax handling, are not tied to specific hardware. These will likely be reimplemented for consumer GPUs over time.

But the full FA-4 kernel, as it exists today, depends on SM100 features. That is not something that can be patched in software.

Bottom line

For local LLM enthusiasts, the takeaway is simple.

RTX Pro 6000 Blackwell is not equivalent to datacenter Blackwell for inference. The SM120 design lacks the hardware needed for FlashAttention-4, and that directly limits performance improvements in modern inference stacks.

If your workload depends on cutting edge attention kernels, architecture matters more than branding.